engine: async NPU encoder compilation on init

[ncylich]

Problem

NPU encoder mlpackages are compiled (compileModelAtURL → .mlmodelc) and loaded synchronously inside Model::init. The CoreML compile is a one-time, per-device cost that is small on CPU/GPU but very large on the ANE path (tens of seconds cold; ~27s for the gemma-4-e2b encoders on an M-series Mac). Because it runs inline in init, it blocks startup for the full compile on a fresh install.

Change

Move the NPU encoder compile/load onto a background thread started early in Model::init:

• start_npu_encoder_loads() launches the three encoder loads via std::async and sets an atomic npu_ready_ flag when done; init returns without waiting.
• has_npu_audio_encoder() / has_npu_vision_encoder() and the source-encode path gate on npu_ready_, so any inference that happens before the NPU is ready cleanly falls back to the existing CPU encoder.
• ~Model joins the load future before the encoders are destroyed.
• Removed the init-time NPU vision soft-token override: the CPU vision-encoder graph already sets the same image soft-token count (verified equal), so it was redundant.

Trade-off

This does not make the compile cheaper — it defers it. init returns fast (~1s instead of ~27s cold); the compile runs in the background while early inferences use the CPU encoder, then the NPU takes over once ready. On warm launches (.mlmodelc cached — the normal case) there is no compile and the NPU is used immediately.

Validation

cactus-engine/tests/test_npu_startup.cpp times init + two completes.

• Cold (no .mlmodelc, ANE forced): init ~1.0s (was ~27s); first completes run on CPU fallback; total process wall time still includes the ~27s compile in the background.
• Warm (.mlmodelc cached): init ~0.6–0.8s, completes uniform, NPU used; correct output, clean exit across repeated runs.